sys/kern/kern_cpu.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2004-2007 Nate Lawson (SDG)
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/bus.h>
  34 #include <sys/cpu.h>
  35 #include <sys/eventhandler.h>
  36 #include <sys/kernel.h>
  37 #include <sys/lock.h>
  38 #include <sys/malloc.h>
  39 #include <sys/module.h>
  40 #include <sys/proc.h>
  41 #include <sys/queue.h>
  42 #include <sys/sbuf.h>
  43 #include <sys/sched.h>
  44 #include <sys/smp.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/systm.h>
  47 #include <sys/sx.h>
  48 #include <sys/timetc.h>
  49 #include <sys/taskqueue.h>
  50
  51 #include "cpufreq_if.h"
  52
  53 /*
  54  * Common CPU frequency glue code.  Drivers for specific hardware can
  55  * attach this interface to allow users to get/set the CPU frequency.
  56  */
  57
  58 /*
  59  * Number of levels we can handle.  Levels are synthesized from settings
  60  * so for M settings and N drivers, there may be M*N levels.
  61  */
  62 #define CF_MAX_LEVELS   256
  63
  64 struct cf_saved_freq {
  65         struct cf_level                 level;
  66         int                             priority;
  67         SLIST_ENTRY(cf_saved_freq)      link;
  68 };
  69
  70 struct cpufreq_softc {
  71         struct sx                       lock;
  72         struct cf_level                 curr_level;
  73         int                             curr_priority;
  74         SLIST_HEAD(, cf_saved_freq)     saved_freq;
  75         struct cf_level_lst             all_levels;
  76         int                             all_count;
  77         int                             max_mhz;
  78         device_t                        dev;
  79         device_t                        cf_drv_dev;
  80         struct sysctl_ctx_list          sysctl_ctx;
  81         struct task                     startup_task;
  82         struct cf_level                 *levels_buf;
  83 };
  84
  85 struct cf_setting_array {
  86         struct cf_setting               sets[MAX_SETTINGS];
  87         int                             count;
  88         TAILQ_ENTRY(cf_setting_array)   link;
  89 };
  90
  91 TAILQ_HEAD(cf_setting_lst, cf_setting_array);
  92
  93 #define CF_MTX_INIT(x)          sx_init((x), "cpufreq lock")
  94 #define CF_MTX_LOCK(x)          sx_xlock((x))
  95 #define CF_MTX_UNLOCK(x)        sx_xunlock((x))
  96 #define CF_MTX_ASSERT(x)        sx_assert((x), SX_XLOCKED)
  97
  98 #define CF_DEBUG(msg...)        do {            \
  99         if (cf_verbose)                         \
 100                 printf("cpufreq: " msg);        \
 101         } while (0)
 102
 103 static int      cpufreq_attach(device_t dev);
 104 static void     cpufreq_startup_task(void *ctx, int pending);
 105 static int      cpufreq_detach(device_t dev);
 106 static int      cf_set_method(device_t dev, const struct cf_level *level,
 107                     int priority);
 108 static int      cf_get_method(device_t dev, struct cf_level *level);
 109 static int      cf_levels_method(device_t dev, struct cf_level *levels,
 110                     int *count);
 111 static int      cpufreq_insert_abs(struct cpufreq_softc *sc,
 112                     struct cf_setting *sets, int count);
 113 static int      cpufreq_expand_set(struct cpufreq_softc *sc,
 114                     struct cf_setting_array *set_arr);
 115 static struct cf_level *cpufreq_dup_set(struct cpufreq_softc *sc,
 116                     struct cf_level *dup, struct cf_setting *set);
 117 static int      cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS);
 118 static int      cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS);
 119 static int      cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS);
 120
 121 static device_method_t cpufreq_methods[] = {
 122         DEVMETHOD(device_probe,         bus_generic_probe),
 123         DEVMETHOD(device_attach,        cpufreq_attach),
 124         DEVMETHOD(device_detach,        cpufreq_detach),
 125
 126         DEVMETHOD(cpufreq_set,          cf_set_method),
 127         DEVMETHOD(cpufreq_get,          cf_get_method),
 128         DEVMETHOD(cpufreq_levels,       cf_levels_method),
 129         {0, 0}
 130 };
 131 static driver_t cpufreq_driver = {
 132         "cpufreq", cpufreq_methods, sizeof(struct cpufreq_softc)
 133 };
 134 static devclass_t cpufreq_dc;
 135 DRIVER_MODULE(cpufreq, cpu, cpufreq_driver, cpufreq_dc, 0, 0);
 136
 137 static int              cf_lowest_freq;
 138 static int              cf_verbose;
 139 static SYSCTL_NODE(_debug, OID_AUTO, cpufreq, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
 140     "cpufreq debugging");
 141 SYSCTL_INT(_debug_cpufreq, OID_AUTO, lowest, CTLFLAG_RWTUN, &cf_lowest_freq, 1,
 142     "Don't provide levels below this frequency.");
 143 SYSCTL_INT(_debug_cpufreq, OID_AUTO, verbose, CTLFLAG_RWTUN, &cf_verbose, 1,
 144     "Print verbose debugging messages");
 145
 146 /*
 147  * This is called as the result of a hardware specific frequency control driver
 148  * calling cpufreq_register. It provides a general interface for system wide
 149  * frequency controls and operates on a per cpu basis.
 150  */
 151 static int
 152 cpufreq_attach(device_t dev)
 153 {
 154         struct cpufreq_softc *sc;
 155         struct pcpu *pc;
 156         device_t parent;
 157         uint64_t rate;
 158
 159         CF_DEBUG("initializing %s\n", device_get_nameunit(dev));
 160         sc = device_get_softc(dev);
 161         parent = device_get_parent(dev);
 162         sc->dev = dev;
 163         sysctl_ctx_init(&sc->sysctl_ctx);
 164         TAILQ_INIT(&sc->all_levels);
 165         CF_MTX_INIT(&sc->lock);
 166         sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 167         SLIST_INIT(&sc->saved_freq);
 168         /* Try to get nominal CPU freq to use it as maximum later if needed */
 169         sc->max_mhz = cpu_get_nominal_mhz(dev);
 170         /* If that fails, try to measure the current rate */
 171         if (sc->max_mhz <= 0) {
 172                 CF_DEBUG("Unable to obtain nominal frequency.\n");
 173                 pc = cpu_get_pcpu(dev);
 174                 if (cpu_est_clockrate(pc->pc_cpuid, &rate) == 0)
 175                         sc->max_mhz = rate / 1000000;
 176                 else
 177                         sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
 178         }
 179
 180         CF_DEBUG("initializing one-time data for %s\n",
 181             device_get_nameunit(dev));
 182         sc->levels_buf = malloc(CF_MAX_LEVELS * sizeof(*sc->levels_buf),
 183             M_DEVBUF, M_WAITOK);
 184         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 185             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 186             OID_AUTO, "freq", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 187             sc, 0, cpufreq_curr_sysctl, "I", "Current CPU frequency");
 188         SYSCTL_ADD_PROC(&sc->sysctl_ctx,
 189             SYSCTL_CHILDREN(device_get_sysctl_tree(parent)),
 190             OID_AUTO, "freq_levels",
 191             CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, sc, 0,
 192             cpufreq_levels_sysctl, "A", "CPU frequency levels");
 193
 194         /*
 195          * Queue a one-shot broadcast that levels have changed.
 196          * It will run once the system has completed booting.
 197          */
 198         TASK_INIT(&sc->startup_task, 0, cpufreq_startup_task, dev);
 199         taskqueue_enqueue(taskqueue_thread, &sc->startup_task);
 200
 201         return (0);
 202 }
 203
 204 /* Handle any work to be done for all drivers that attached during boot. */
 205 static void
 206 cpufreq_startup_task(void *ctx, int pending)
 207 {
 208
 209         cpufreq_settings_changed((device_t)ctx);
 210 }
 211
 212 static int
 213 cpufreq_detach(device_t dev)
 214 {
 215         struct cpufreq_softc *sc;
 216         struct cf_saved_freq *saved_freq;
 217
 218         CF_DEBUG("shutdown %s\n", device_get_nameunit(dev));
 219         sc = device_get_softc(dev);
 220         sysctl_ctx_free(&sc->sysctl_ctx);
 221
 222         while ((saved_freq = SLIST_FIRST(&sc->saved_freq)) != NULL) {
 223                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 224                 free(saved_freq, M_TEMP);
 225         }
 226
 227         free(sc->levels_buf, M_DEVBUF);
 228
 229         return (0);
 230 }
 231
 232 static int
 233 cf_set_method(device_t dev, const struct cf_level *level, int priority)
 234 {
 235         struct cpufreq_softc *sc;
 236         const struct cf_setting *set;
 237         struct cf_saved_freq *saved_freq, *curr_freq;
 238         struct pcpu *pc;
 239         int error, i;
 240         u_char pri;
 241
 242         sc = device_get_softc(dev);
 243         error = 0;
 244         set = NULL;
 245         saved_freq = NULL;
 246
 247         /* We are going to change levels so notify the pre-change handler. */
 248         EVENTHANDLER_INVOKE(cpufreq_pre_change, level, &error);
 249         if (error != 0) {
 250                 EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 251                 return (error);
 252         }
 253
 254         CF_MTX_LOCK(&sc->lock);
 255
 256 #ifdef SMP
 257 #ifdef EARLY_AP_STARTUP
 258         MPASS(mp_ncpus == 1 || smp_started);
 259 #else
 260         /*
 261          * If still booting and secondary CPUs not started yet, don't allow
 262          * changing the frequency until they're online.  This is because we
 263          * can't switch to them using sched_bind() and thus we'd only be
 264          * switching the main CPU.  XXXTODO: Need to think more about how to
 265          * handle having different CPUs at different frequencies.
 266          */
 267         if (mp_ncpus > 1 && !smp_started) {
 268                 device_printf(dev, "rejecting change, SMP not started yet\n");
 269                 error = ENXIO;
 270                 goto out;
 271         }
 272 #endif
 273 #endif /* SMP */
 274
 275         /*
 276          * If the requested level has a lower priority, don't allow
 277          * the new level right now.
 278          */
 279         if (priority < sc->curr_priority) {
 280                 CF_DEBUG("ignoring, curr prio %d less than %d\n", priority,
 281                     sc->curr_priority);
 282                 error = EPERM;
 283                 goto out;
 284         }
 285
 286         /*
 287          * If the caller didn't specify a level and one is saved, prepare to
 288          * restore the saved level.  If none has been saved, return an error.
 289          */
 290         if (level == NULL) {
 291                 saved_freq = SLIST_FIRST(&sc->saved_freq);
 292                 if (saved_freq == NULL) {
 293                         CF_DEBUG("NULL level, no saved level\n");
 294                         error = ENXIO;
 295                         goto out;
 296                 }
 297                 level = &saved_freq->level;
 298                 priority = saved_freq->priority;
 299                 CF_DEBUG("restoring saved level, freq %d prio %d\n",
 300                     level->total_set.freq, priority);
 301         }
 302
 303         /* Reject levels that are below our specified threshold. */
 304         if (level->total_set.freq < cf_lowest_freq) {
 305                 CF_DEBUG("rejecting freq %d, less than %d limit\n",
 306                     level->total_set.freq, cf_lowest_freq);
 307                 error = EINVAL;
 308                 goto out;
 309         }
 310
 311         /* If already at this level, just return. */
 312         if (sc->curr_level.total_set.freq == level->total_set.freq) {
 313                 CF_DEBUG("skipping freq %d, same as current level %d\n",
 314                     level->total_set.freq, sc->curr_level.total_set.freq);
 315                 goto skip;
 316         }
 317
 318         /* First, set the absolute frequency via its driver. */
 319         set = &level->abs_set;
 320         if (set->dev) {
 321                 if (!device_is_attached(set->dev)) {
 322                         error = ENXIO;
 323                         goto out;
 324                 }
 325
 326                 /* Bind to the target CPU before switching. */
 327                 pc = cpu_get_pcpu(set->dev);
 328                 thread_lock(curthread);
 329                 pri = curthread->td_priority;
 330                 sched_prio(curthread, PRI_MIN);
 331                 sched_bind(curthread, pc->pc_cpuid);
 332                 thread_unlock(curthread);
 333                 CF_DEBUG("setting abs freq %d on %s (cpu %d)\n", set->freq,
 334                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 335                 error = CPUFREQ_DRV_SET(set->dev, set);
 336                 thread_lock(curthread);
 337                 sched_unbind(curthread);
 338                 sched_prio(curthread, pri);
 339                 thread_unlock(curthread);
 340                 if (error) {
 341                         goto out;
 342                 }
 343         }
 344
 345         /* Next, set any/all relative frequencies via their drivers. */
 346         for (i = 0; i < level->rel_count; i++) {
 347                 set = &level->rel_set[i];
 348                 if (!device_is_attached(set->dev)) {
 349                         error = ENXIO;
 350                         goto out;
 351                 }
 352
 353                 /* Bind to the target CPU before switching. */
 354                 pc = cpu_get_pcpu(set->dev);
 355                 thread_lock(curthread);
 356                 pri = curthread->td_priority;
 357                 sched_prio(curthread, PRI_MIN);
 358                 sched_bind(curthread, pc->pc_cpuid);
 359                 thread_unlock(curthread);
 360                 CF_DEBUG("setting rel freq %d on %s (cpu %d)\n", set->freq,
 361                     device_get_nameunit(set->dev), PCPU_GET(cpuid));
 362                 error = CPUFREQ_DRV_SET(set->dev, set);
 363                 thread_lock(curthread);
 364                 sched_unbind(curthread);
 365                 sched_prio(curthread, pri);
 366                 thread_unlock(curthread);
 367                 if (error) {
 368                         /* XXX Back out any successful setting? */
 369                         goto out;
 370                 }
 371         }
 372
 373 skip:
 374         /*
 375          * Before recording the current level, check if we're going to a
 376          * higher priority.  If so, save the previous level and priority.
 377          */
 378         if (sc->curr_level.total_set.freq != CPUFREQ_VAL_UNKNOWN &&
 379             priority > sc->curr_priority) {
 380                 CF_DEBUG("saving level, freq %d prio %d\n",
 381                     sc->curr_level.total_set.freq, sc->curr_priority);
 382                 curr_freq = malloc(sizeof(*curr_freq), M_TEMP, M_NOWAIT);
 383                 if (curr_freq == NULL) {
 384                         error = ENOMEM;
 385                         goto out;
 386                 }
 387                 curr_freq->level = sc->curr_level;
 388                 curr_freq->priority = sc->curr_priority;
 389                 SLIST_INSERT_HEAD(&sc->saved_freq, curr_freq, link);
 390         }
 391         sc->curr_level = *level;
 392         sc->curr_priority = priority;
 393
 394         /* If we were restoring a saved state, reset it to "unused". */
 395         if (saved_freq != NULL) {
 396                 CF_DEBUG("resetting saved level\n");
 397                 sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN;
 398                 SLIST_REMOVE_HEAD(&sc->saved_freq, link);
 399                 free(saved_freq, M_TEMP);
 400         }
 401
 402 out:
 403         CF_MTX_UNLOCK(&sc->lock);
 404
 405         /*
 406          * We changed levels (or attempted to) so notify the post-change
 407          * handler of new frequency or error.
 408          */
 409         EVENTHANDLER_INVOKE(cpufreq_post_change, level, error);
 410         if (error && set)
 411                 device_printf(set->dev, "set freq failed, err %d\n", error);
 412
 413         return (error);
 414 }
 415
 416 static int
 417 cpufreq_get_frequency(device_t dev)
 418 {
 419         struct cf_setting set;
 420
 421         if (CPUFREQ_DRV_GET(dev, &set) != 0)
 422                 return (-1);
 423
 424         return (set.freq);
 425 }
 426
 427 /* Returns the index into *levels with the match */
 428 static int
 429 cpufreq_get_level(device_t dev, struct cf_level *levels, int count)
 430 {
 431         int i, freq;
 432
 433         if ((freq = cpufreq_get_frequency(dev)) < 0)
 434                 return (-1);
 435         for (i = 0; i < count; i++)
 436                 if (freq == levels[i].total_set.freq)
 437                         return (i);
 438
 439         return (-1);
 440 }
 441
 442 /*
 443  * Used by the cpufreq core, this function will populate *level with the current
 444  * frequency as either determined by a cached value sc->curr_level, or in the
 445  * case the lower level driver has set the CPUFREQ_FLAG_UNCACHED flag, it will
 446  * obtain the frequency from the driver itself.
 447  */
 448 static int
 449 cf_get_method(device_t dev, struct cf_level *level)
 450 {
 451         struct cpufreq_softc *sc;
 452         struct cf_level *levels;
 453         struct cf_setting *curr_set;
 454         struct pcpu *pc;
 455         int bdiff, count, diff, error, i, type;
 456         uint64_t rate;
 457
 458         sc = device_get_softc(dev);
 459         error = 0;
 460         levels = NULL;
 461
 462         /*
 463          * If we already know the current frequency, and the driver didn't ask
 464          * for uncached usage, we're done.
 465          */
 466         CF_MTX_LOCK(&sc->lock);
 467         curr_set = &sc->curr_level.total_set;
 468         error = CPUFREQ_DRV_TYPE(sc->cf_drv_dev, &type);
 469         if (error == 0 && (type & CPUFREQ_FLAG_UNCACHED)) {
 470                 struct cf_setting set;
 471
 472                 /*
 473                  * If the driver wants to always report back the real frequency,
 474                  * first try the driver and if that fails, fall back to
 475                  * estimating.
 476                  */
 477                 if (CPUFREQ_DRV_GET(sc->cf_drv_dev, &set) == 0) {
 478                         sc->curr_level.total_set = set;
 479                         CF_DEBUG("get returning immediate freq %d\n",
 480                             curr_set->freq);
 481                         goto out;
 482                 }
 483         } else if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 484                 CF_DEBUG("get returning known freq %d\n", curr_set->freq);
 485                 error = 0;
 486                 goto out;
 487         }
 488         CF_MTX_UNLOCK(&sc->lock);
 489
 490         /*
 491          * We need to figure out the current level.  Loop through every
 492          * driver, getting the current setting.  Then, attempt to get a best
 493          * match of settings against each level.
 494          */
 495         count = CF_MAX_LEVELS;
 496         levels = malloc(count * sizeof(*levels), M_TEMP, M_NOWAIT);
 497         if (levels == NULL)
 498                 return (ENOMEM);
 499         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
 500         if (error) {
 501                 if (error == E2BIG)
 502                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
 503                 free(levels, M_TEMP);
 504                 return (error);
 505         }
 506
 507         /*
 508          * Reacquire the lock and search for the given level.
 509          *
 510          * XXX Note: this is not quite right since we really need to go
 511          * through each level and compare both absolute and relative
 512          * settings for each driver in the system before making a match.
 513          * The estimation code below catches this case though.
 514          */
 515         CF_MTX_LOCK(&sc->lock);
 516         i = cpufreq_get_level(sc->cf_drv_dev, levels, count);
 517         if (i >= 0)
 518                 sc->curr_level = levels[i];
 519         else
 520                 CF_DEBUG("Couldn't find supported level for %s\n",
 521                     device_get_nameunit(sc->cf_drv_dev));
 522
 523         if (curr_set->freq != CPUFREQ_VAL_UNKNOWN) {
 524                 CF_DEBUG("get matched freq %d from drivers\n", curr_set->freq);
 525                 goto out;
 526         }
 527
 528         /*
 529          * We couldn't find an exact match, so attempt to estimate and then
 530          * match against a level.
 531          */
 532         pc = cpu_get_pcpu(dev);
 533         if (pc == NULL) {
 534                 error = ENXIO;
 535                 goto out;
 536         }
 537         cpu_est_clockrate(pc->pc_cpuid, &rate);
 538         rate /= 1000000;
 539         bdiff = 1 << 30;
 540         for (i = 0; i < count; i++) {
 541                 diff = abs(levels[i].total_set.freq - rate);
 542                 if (diff < bdiff) {
 543                         bdiff = diff;
 544                         sc->curr_level = levels[i];
 545                 }
 546         }
 547         CF_DEBUG("get estimated freq %d\n", curr_set->freq);
 548
 549 out:
 550         if (error == 0)
 551                 *level = sc->curr_level;
 552
 553         CF_MTX_UNLOCK(&sc->lock);
 554         if (levels)
 555                 free(levels, M_TEMP);
 556         return (error);
 557 }
 558
 559 /*
 560  * Either directly obtain settings from the cpufreq driver, or build a list of
 561  * relative settings to be integrated later against an absolute max.
 562  */
 563 static int
 564 cpufreq_add_levels(device_t cf_dev, struct cf_setting_lst *rel_sets)
 565 {
 566         struct cf_setting_array *set_arr;
 567         struct cf_setting *sets;
 568         device_t dev;
 569         struct cpufreq_softc *sc;
 570         int type, set_count, error;
 571
 572         sc = device_get_softc(cf_dev);
 573         dev = sc->cf_drv_dev;
 574
 575         /* Skip devices that aren't ready. */
 576         if (!device_is_attached(cf_dev))
 577                 return (0);
 578
 579         /*
 580          * Get settings, skipping drivers that offer no settings or
 581          * provide settings for informational purposes only.
 582          */
 583         error = CPUFREQ_DRV_TYPE(dev, &type);
 584         if (error != 0 || (type & CPUFREQ_FLAG_INFO_ONLY)) {
 585                 if (error == 0) {
 586                         CF_DEBUG("skipping info-only driver %s\n",
 587                             device_get_nameunit(cf_dev));
 588                 }
 589                 return (error);
 590         }
 591
 592         sets = malloc(MAX_SETTINGS * sizeof(*sets), M_TEMP, M_NOWAIT);
 593         if (sets == NULL)
 594                 return (ENOMEM);
 595
 596         set_count = MAX_SETTINGS;
 597         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
 598         if (error != 0 || set_count == 0)
 599                 goto out;
 600
 601         /* Add the settings to our absolute/relative lists. */
 602         switch (type & CPUFREQ_TYPE_MASK) {
 603         case CPUFREQ_TYPE_ABSOLUTE:
 604                 error = cpufreq_insert_abs(sc, sets, set_count);
 605                 break;
 606         case CPUFREQ_TYPE_RELATIVE:
 607                 CF_DEBUG("adding %d relative settings\n", set_count);
 608                 set_arr = malloc(sizeof(*set_arr), M_TEMP, M_NOWAIT);
 609                 if (set_arr == NULL) {
 610                         error = ENOMEM;
 611                         goto out;
 612                 }
 613                 bcopy(sets, set_arr->sets, set_count * sizeof(*sets));
 614                 set_arr->count = set_count;
 615                 TAILQ_INSERT_TAIL(rel_sets, set_arr, link);
 616                 break;
 617         default:
 618                 error = EINVAL;
 619         }
 620
 621 out:
 622         free(sets, M_TEMP);
 623         return (error);
 624 }
 625
 626 static int
 627 cf_levels_method(device_t dev, struct cf_level *levels, int *count)
 628 {
 629         struct cf_setting_array *set_arr;
 630         struct cf_setting_lst rel_sets;
 631         struct cpufreq_softc *sc;
 632         struct cf_level *lev;
 633         struct pcpu *pc;
 634         int error, i;
 635         uint64_t rate;
 636
 637         if (levels == NULL || count == NULL)
 638                 return (EINVAL);
 639
 640         TAILQ_INIT(&rel_sets);
 641         sc = device_get_softc(dev);
 642
 643         CF_MTX_LOCK(&sc->lock);
 644         error = cpufreq_add_levels(sc->dev, &rel_sets);
 645         if (error)
 646                 goto out;
 647
 648         /*
 649          * If there are no absolute levels, create a fake one at 100%.  We
 650          * then cache the clockrate for later use as our base frequency.
 651          */
 652         if (TAILQ_EMPTY(&sc->all_levels)) {
 653                 struct cf_setting set;
 654
 655                 CF_DEBUG("No absolute levels returned by driver\n");
 656
 657                 if (sc->max_mhz == CPUFREQ_VAL_UNKNOWN) {
 658                         sc->max_mhz = cpu_get_nominal_mhz(dev);
 659                         /*
 660                          * If the CPU can't report a rate for 100%, hope
 661                          * the CPU is running at its nominal rate right now,
 662                          * and use that instead.
 663                          */
 664                         if (sc->max_mhz <= 0) {
 665                                 pc = cpu_get_pcpu(dev);
 666                                 cpu_est_clockrate(pc->pc_cpuid, &rate);
 667                                 sc->max_mhz = rate / 1000000;
 668                         }
 669                 }
 670                 memset(&set, CPUFREQ_VAL_UNKNOWN, sizeof(set));
 671                 set.freq = sc->max_mhz;
 672                 set.dev = NULL;
 673                 error = cpufreq_insert_abs(sc, &set, 1);
 674                 if (error)
 675                         goto out;
 676         }
 677
 678         /* Create a combined list of absolute + relative levels. */
 679         TAILQ_FOREACH(set_arr, &rel_sets, link)
 680                 cpufreq_expand_set(sc, set_arr);
 681
 682         /* If the caller doesn't have enough space, return the actual count. */
 683         if (sc->all_count > *count) {
 684                 *count = sc->all_count;
 685                 error = E2BIG;
 686                 goto out;
 687         }
 688
 689         /* Finally, output the list of levels. */
 690         i = 0;
 691         TAILQ_FOREACH(lev, &sc->all_levels, link) {
 692
 693                 /* Skip levels that have a frequency that is too low. */
 694                 if (lev->total_set.freq < cf_lowest_freq) {
 695                         sc->all_count--;
 696                         continue;
 697                 }
 698
 699                 levels[i] = *lev;
 700                 i++;
 701         }
 702         *count = sc->all_count;
 703         error = 0;
 704
 705 out:
 706         /* Clear all levels since we regenerate them each time. */
 707         while ((lev = TAILQ_FIRST(&sc->all_levels)) != NULL) {
 708                 TAILQ_REMOVE(&sc->all_levels, lev, link);
 709                 free(lev, M_TEMP);
 710         }
 711         sc->all_count = 0;
 712
 713         CF_MTX_UNLOCK(&sc->lock);
 714         while ((set_arr = TAILQ_FIRST(&rel_sets)) != NULL) {
 715                 TAILQ_REMOVE(&rel_sets, set_arr, link);
 716                 free(set_arr, M_TEMP);
 717         }
 718         return (error);
 719 }
 720
 721 /*
 722  * Create levels for an array of absolute settings and insert them in
 723  * sorted order in the specified list.
 724  */
 725 static int
 726 cpufreq_insert_abs(struct cpufreq_softc *sc, struct cf_setting *sets,
 727     int count)
 728 {
 729         struct cf_level_lst *list;
 730         struct cf_level *level, *search;
 731         int i, inserted;
 732
 733         CF_MTX_ASSERT(&sc->lock);
 734
 735         list = &sc->all_levels;
 736         for (i = 0; i < count; i++) {
 737                 level = malloc(sizeof(*level), M_TEMP, M_NOWAIT | M_ZERO);
 738                 if (level == NULL)
 739                         return (ENOMEM);
 740                 level->abs_set = sets[i];
 741                 level->total_set = sets[i];
 742                 level->total_set.dev = NULL;
 743                 sc->all_count++;
 744                 inserted = 0;
 745
 746                 if (TAILQ_EMPTY(list)) {
 747                         CF_DEBUG("adding abs setting %d at head\n",
 748                             sets[i].freq);
 749                         TAILQ_INSERT_HEAD(list, level, link);
 750                         continue;
 751                 }
 752
 753                 TAILQ_FOREACH_REVERSE(search, list, cf_level_lst, link)
 754                         if (sets[i].freq <= search->total_set.freq) {
 755                                 CF_DEBUG("adding abs setting %d after %d\n",
 756                                     sets[i].freq, search->total_set.freq);
 757                                 TAILQ_INSERT_AFTER(list, search, level, link);
 758                                 inserted = 1;
 759                                 break;
 760                         }
 761
 762                 if (inserted == 0) {
 763                         TAILQ_FOREACH(search, list, link)
 764                                 if (sets[i].freq >= search->total_set.freq) {
 765                                         CF_DEBUG("adding abs setting %d before %d\n",
 766                                             sets[i].freq, search->total_set.freq);
 767                                         TAILQ_INSERT_BEFORE(search, level, link);
 768                                         break;
 769                                 }
 770                 }
 771         }
 772
 773         return (0);
 774 }
 775
 776 /*
 777  * Expand a group of relative settings, creating derived levels from them.
 778  */
 779 static int
 780 cpufreq_expand_set(struct cpufreq_softc *sc, struct cf_setting_array *set_arr)
 781 {
 782         struct cf_level *fill, *search;
 783         struct cf_setting *set;
 784         int i;
 785
 786         CF_MTX_ASSERT(&sc->lock);
 787
 788         /*
 789          * Walk the set of all existing levels in reverse.  This is so we
 790          * create derived states from the lowest absolute settings first
 791          * and discard duplicates created from higher absolute settings.
 792          * For instance, a level of 50 Mhz derived from 100 Mhz + 50% is
 793          * preferable to 200 Mhz + 25% because absolute settings are more
 794          * efficient since they often change the voltage as well.
 795          */
 796         TAILQ_FOREACH_REVERSE(search, &sc->all_levels, cf_level_lst, link) {
 797                 /* Add each setting to the level, duplicating if necessary. */
 798                 for (i = 0; i < set_arr->count; i++) {
 799                         set = &set_arr->sets[i];
 800
 801                         /*
 802                          * If this setting is less than 100%, split the level
 803                          * into two and add this setting to the new level.
 804                          */
 805                         fill = search;
 806                         if (set->freq < 10000) {
 807                                 fill = cpufreq_dup_set(sc, search, set);
 808
 809                                 /*
 810                                  * The new level was a duplicate of an existing
 811                                  * level or its absolute setting is too high
 812                                  * so we freed it.  For example, we discard a
 813                                  * derived level of 1000 MHz/25% if a level
 814                                  * of 500 MHz/100% already exists.
 815                                  */
 816                                 if (fill == NULL)
 817                                         break;
 818                         }
 819
 820                         /* Add this setting to the existing or new level. */
 821                         KASSERT(fill->rel_count < MAX_SETTINGS,
 822                             ("cpufreq: too many relative drivers (%d)",
 823                             MAX_SETTINGS));
 824                         fill->rel_set[fill->rel_count] = *set;
 825                         fill->rel_count++;
 826                         CF_DEBUG(
 827                         "expand set added rel setting %d%% to %d level\n",
 828                             set->freq / 100, fill->total_set.freq);
 829                 }
 830         }
 831
 832         return (0);
 833 }
 834
 835 static struct cf_level *
 836 cpufreq_dup_set(struct cpufreq_softc *sc, struct cf_level *dup,
 837     struct cf_setting *set)
 838 {
 839         struct cf_level_lst *list;
 840         struct cf_level *fill, *itr;
 841         struct cf_setting *fill_set, *itr_set;
 842         int i;
 843
 844         CF_MTX_ASSERT(&sc->lock);
 845
 846         /*
 847          * Create a new level, copy it from the old one, and update the
 848          * total frequency and power by the percentage specified in the
 849          * relative setting.
 850          */
 851         fill = malloc(sizeof(*fill), M_TEMP, M_NOWAIT);
 852         if (fill == NULL)
 853                 return (NULL);
 854         *fill = *dup;
 855         fill_set = &fill->total_set;
 856         fill_set->freq =
 857             ((uint64_t)fill_set->freq * set->freq) / 10000;
 858         if (fill_set->power != CPUFREQ_VAL_UNKNOWN) {
 859                 fill_set->power = ((uint64_t)fill_set->power * set->freq)
 860                     / 10000;
 861         }
 862         if (set->lat != CPUFREQ_VAL_UNKNOWN) {
 863                 if (fill_set->lat != CPUFREQ_VAL_UNKNOWN)
 864                         fill_set->lat += set->lat;
 865                 else
 866                         fill_set->lat = set->lat;
 867         }
 868         CF_DEBUG("dup set considering derived setting %d\n", fill_set->freq);
 869
 870         /*
 871          * If we copied an old level that we already modified (say, at 100%),
 872          * we need to remove that setting before adding this one.  Since we
 873          * process each setting array in order, we know any settings for this
 874          * driver will be found at the end.
 875          */
 876         for (i = fill->rel_count; i != 0; i--) {
 877                 if (fill->rel_set[i - 1].dev != set->dev)
 878                         break;
 879                 CF_DEBUG("removed last relative driver: %s\n",
 880                     device_get_nameunit(set->dev));
 881                 fill->rel_count--;
 882         }
 883
 884         /*
 885          * Insert the new level in sorted order.  If it is a duplicate of an
 886          * existing level (1) or has an absolute setting higher than the
 887          * existing level (2), do not add it.  We can do this since any such
 888          * level is guaranteed use less power.  For example (1), a level with
 889          * one absolute setting of 800 Mhz uses less power than one composed
 890          * of an absolute setting of 1600 Mhz and a relative setting at 50%.
 891          * Also for example (2), a level of 800 Mhz/75% is preferable to
 892          * 1600 Mhz/25% even though the latter has a lower total frequency.
 893          */
 894         list = &sc->all_levels;
 895         KASSERT(!TAILQ_EMPTY(list), ("all levels list empty in dup set"));
 896         TAILQ_FOREACH_REVERSE(itr, list, cf_level_lst, link) {
 897                 itr_set = &itr->total_set;
 898                 if (CPUFREQ_CMP(fill_set->freq, itr_set->freq)) {
 899                         CF_DEBUG("dup set rejecting %d (dupe)\n",
 900                             fill_set->freq);
 901                         itr = NULL;
 902                         break;
 903                 } else if (fill_set->freq < itr_set->freq) {
 904                         if (fill->abs_set.freq <= itr->abs_set.freq) {
 905                                 CF_DEBUG(
 906                         "dup done, inserting new level %d after %d\n",
 907                                     fill_set->freq, itr_set->freq);
 908                                 TAILQ_INSERT_AFTER(list, itr, fill, link);
 909                                 sc->all_count++;
 910                         } else {
 911                                 CF_DEBUG("dup set rejecting %d (abs too big)\n",
 912                                     fill_set->freq);
 913                                 itr = NULL;
 914                         }
 915                         break;
 916                 }
 917         }
 918
 919         /* We didn't find a good place for this new level so free it. */
 920         if (itr == NULL) {
 921                 CF_DEBUG("dup set freeing new level %d (not optimal)\n",
 922                     fill_set->freq);
 923                 free(fill, M_TEMP);
 924                 fill = NULL;
 925         }
 926
 927         return (fill);
 928 }
 929
 930 static int
 931 cpufreq_curr_sysctl(SYSCTL_HANDLER_ARGS)
 932 {
 933         struct cpufreq_softc *sc;
 934         struct cf_level *levels;
 935         int best, count, diff, bdiff, devcount, error, freq, i, n;
 936         device_t *devs;
 937
 938         devs = NULL;
 939         sc = oidp->oid_arg1;
 940         levels = sc->levels_buf;
 941
 942         error = CPUFREQ_GET(sc->dev, &levels[0]);
 943         if (error)
 944                 goto out;
 945         freq = levels[0].total_set.freq;
 946         error = sysctl_handle_int(oidp, &freq, 0, req);
 947         if (error != 0 || req->newptr == NULL)
 948                 goto out;
 949
 950         /*
 951          * While we only call cpufreq_get() on one device (assuming all
 952          * CPUs have equal levels), we call cpufreq_set() on all CPUs.
 953          * This is needed for some MP systems.
 954          */
 955         error = devclass_get_devices(cpufreq_dc, &devs, &devcount);
 956         if (error)
 957                 goto out;
 958         for (n = 0; n < devcount; n++) {
 959                 count = CF_MAX_LEVELS;
 960                 error = CPUFREQ_LEVELS(devs[n], levels, &count);
 961                 if (error) {
 962                         if (error == E2BIG)
 963                                 printf(
 964                         "cpufreq: need to increase CF_MAX_LEVELS\n");
 965                         break;
 966                 }
 967                 best = 0;
 968                 bdiff = 1 << 30;
 969                 for (i = 0; i < count; i++) {
 970                         diff = abs(levels[i].total_set.freq - freq);
 971                         if (diff < bdiff) {
 972                                 bdiff = diff;
 973                                 best = i;
 974                         }
 975                 }
 976                 error = CPUFREQ_SET(devs[n], &levels[best], CPUFREQ_PRIO_USER);
 977         }
 978
 979 out:
 980         if (devs)
 981                 free(devs, M_TEMP);
 982         return (error);
 983 }
 984
 985 static int
 986 cpufreq_levels_sysctl(SYSCTL_HANDLER_ARGS)
 987 {
 988         struct cpufreq_softc *sc;
 989         struct cf_level *levels;
 990         struct cf_setting *set;
 991         struct sbuf sb;
 992         int count, error, i;
 993
 994         sc = oidp->oid_arg1;
 995         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
 996
 997         /* Get settings from the device and generate the output string. */
 998         count = CF_MAX_LEVELS;
 999         levels = sc->levels_buf;
1000         if (levels == NULL) {
1001                 sbuf_delete(&sb);
1002                 return (ENOMEM);
1003         }
1004         error = CPUFREQ_LEVELS(sc->dev, levels, &count);
1005         if (error) {
1006                 if (error == E2BIG)
1007                         printf("cpufreq: need to increase CF_MAX_LEVELS\n");
1008                 goto out;
1009         }
1010         if (count) {
1011                 for (i = 0; i < count; i++) {
1012                         set = &levels[i].total_set;
1013                         sbuf_printf(&sb, "%d/%d ", set->freq, set->power);
1014                 }
1015         } else
1016                 sbuf_cpy(&sb, "0");
1017         sbuf_trim(&sb);
1018         sbuf_finish(&sb);
1019         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1020
1021 out:
1022         sbuf_delete(&sb);
1023         return (error);
1024 }
1025
1026 static int
1027 cpufreq_settings_sysctl(SYSCTL_HANDLER_ARGS)
1028 {
1029         device_t dev;
1030         struct cf_setting *sets;
1031         struct sbuf sb;
1032         int error, i, set_count;
1033
1034         dev = oidp->oid_arg1;
1035         sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
1036
1037         /* Get settings from the device and generate the output string. */
1038         set_count = MAX_SETTINGS;
1039         sets = malloc(set_count * sizeof(*sets), M_TEMP, M_NOWAIT);
1040         if (sets == NULL) {
1041                 sbuf_delete(&sb);
1042                 return (ENOMEM);
1043         }
1044         error = CPUFREQ_DRV_SETTINGS(dev, sets, &set_count);
1045         if (error)
1046                 goto out;
1047         if (set_count) {
1048                 for (i = 0; i < set_count; i++)
1049                         sbuf_printf(&sb, "%d/%d ", sets[i].freq, sets[i].power);
1050         } else
1051                 sbuf_cpy(&sb, "0");
1052         sbuf_trim(&sb);
1053         sbuf_finish(&sb);
1054         error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
1055
1056 out:
1057         free(sets, M_TEMP);
1058         sbuf_delete(&sb);
1059         return (error);
1060 }
1061
1062 static void
1063 cpufreq_add_freq_driver_sysctl(device_t cf_dev)
1064 {
1065         struct cpufreq_softc *sc;
1066
1067         sc = device_get_softc(cf_dev);
1068         SYSCTL_ADD_CONST_STRING(&sc->sysctl_ctx,
1069             SYSCTL_CHILDREN(device_get_sysctl_tree(cf_dev)), OID_AUTO,
1070             "freq_driver", CTLFLAG_RD, device_get_nameunit(sc->cf_drv_dev),
1071             "cpufreq driver used by this cpu");
1072 }
1073
1074 int
1075 cpufreq_register(device_t dev)
1076 {
1077         struct cpufreq_softc *sc;
1078         device_t cf_dev, cpu_dev;
1079         int error;
1080
1081         /* Add a sysctl to get each driver's settings separately. */
1082         SYSCTL_ADD_PROC(device_get_sysctl_ctx(dev),
1083             SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
1084             OID_AUTO, "freq_settings",
1085             CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, dev, 0,
1086             cpufreq_settings_sysctl, "A", "CPU frequency driver settings");
1087
1088         /*
1089          * Add only one cpufreq device to each CPU.  Currently, all CPUs
1090          * must offer the same levels and be switched at the same time.
1091          */
1092         cpu_dev = device_get_parent(dev);
1093         if ((cf_dev = device_find_child(cpu_dev, "cpufreq", -1))) {
1094                 sc = device_get_softc(cf_dev);
1095                 sc->max_mhz = CPUFREQ_VAL_UNKNOWN;
1096                 MPASS(sc->cf_drv_dev != NULL);
1097                 return (0);
1098         }
1099
1100         /* Add the child device and possibly sysctls. */
1101         cf_dev = BUS_ADD_CHILD(cpu_dev, 0, "cpufreq", -1);
1102         if (cf_dev == NULL)
1103                 return (ENOMEM);
1104         device_quiet(cf_dev);
1105
1106         error = device_probe_and_attach(cf_dev);
1107         if (error)
1108                 return (error);
1109
1110         sc = device_get_softc(cf_dev);
1111         sc->cf_drv_dev = dev;
1112         cpufreq_add_freq_driver_sysctl(cf_dev);
1113         return (error);
1114 }
1115
1116 int
1117 cpufreq_unregister(device_t dev)
1118 {
1119         device_t cf_dev;
1120         struct cpufreq_softc *sc;
1121
1122         /*
1123          * If this is the last cpufreq child device, remove the control
1124          * device as well.  We identify cpufreq children by calling a method
1125          * they support.
1126          */
1127         cf_dev = device_find_child(device_get_parent(dev), "cpufreq", -1);
1128         if (cf_dev == NULL) {
1129                 device_printf(dev,
1130         "warning: cpufreq_unregister called with no cpufreq device active\n");
1131                 return (0);
1132         }
1133         sc = device_get_softc(cf_dev);
1134         MPASS(sc->cf_drv_dev == dev);
1135         device_delete_child(device_get_parent(cf_dev), cf_dev);
1136
1137         return (0);
1138 }
1139
1140 int
1141 cpufreq_settings_changed(device_t dev)
1142 {
1143
1144         EVENTHANDLER_INVOKE(cpufreq_levels_changed,
1145             device_get_unit(device_get_parent(dev)));
1146         return (0);
1147 }